AM68A: non-box nms using TIDL

Roman Milishchuk

Part Number: AM68A
Other Parts Discussed in Thread: AM68

We are currently experimenting with deep-learning-based keypoint extractors and matchers on AM68/TDA4 and need to perform nms over score grid instead of bbox output of yolo. Do you think it is possible to perform tidl acceleration for it?

here are reference implementations:

https://github.com/magicleap/SuperGluePretrainedNetwork/blob/master/models/superpoint.py#L47
https://github.com/magicleap/SuperPointPretrainedNetwork/blob/master/demo_superpoint.py#L151

over 1 year ago

0 Tanmay Patil over 1 year ago

TI__Mastermind 38920 points

Hi,

Its a TI-India holiday today. Please expect a delay in response.

Regards,
Tanmay

0 Pratik Kedar over 1 year ago in reply to Tanmay Patil

TI__Mastermind 24041 points

Hi,

For post processing specific optimization we have supported standard ones, as part of meta architectures.

We recommend to please do diligence of it here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_06_00/docs/tidl_fsg_od_meta_arch.md

And these are ones which are supported at TIDL end, anything which is different seems to fall under unsupported category and that particular operator or operators will get delegated to arm.

0 Roman Milishchuk over 1 year ago in reply to Pratik Kedar

Prodigy 200 points

Is there any way to add custom operator to tidl-tools ?

0 Pratik Kedar over 1 year ago in reply to Roman Milishchuk

TI__Mastermind 24041 points

Hi,

Roman Milishchuk said:
Is there any way to add custom operator to tidl-tools ?

Can you please connect to local FAE and he will get back to us.

Thanks

Processors

Processors forum

AM68A: non-box nms using TIDL